3.71 GB
15 files
Updated 5 months ago
README.md

Specialist LoRAs

Narrow, single-purpose edit LoRAs trained on top of the NK2E generalist (see the main README). They load and run exactly like the generalist edit LoRAs — same ComfyUI node setup, source image as the reference — but each responds to one specific trigger instruction instead of general edits.

Experimental and unofficial. Results are rough and inconsistent. Not affiliated with or endorsed by Krea.

Use (ComfyUI)

  1. Download the .safetensors into your ComfyUI models/loras/.
  2. Load the NK2E example workflow (node pack: https://github.com/Nynxz/ComfyUI-NK2E).
  3. Set the source image as the reference and use the specialist's trigger as the text instruction.

Specialists are low-rank and concept-specific, so they usually want a higher strength than the generalist — start near 1.0 and back off if the edit is too aggressive.

Available specialists

Name File Base Rank Trigger Strength
Invisible Man NK2E-InvisibleMan-v0.1.safetensors v0.3 8 make person invisible ~1.0

NK2E Invisible Man

Removes the person from the source image as if they had turned invisible — clothing, props and the surrounding scene stay, the body does not.

  • File: NK2E-InvisibleMan-v0.1.safetensors
  • Trigger: make person invisible
  • Base: trained on top of the v0.3 generalist (train/v0.3/NK2E-v0.3-train.safetensors)
  • Rank: 8
  • Recommended strength: ~1.0

Feed a photo of a person as the source image and use the trigger as the instruction. Works best with a clear single subject; busy scenes and multiple people are less reliable.

Training your own

These were trained by pointing the toolkit's base_lora at the latest generalist native weights and fine-tuning on a small, focused dataset. See the training toolkit: https://github.com/Nynxz/NK2E.

Total size
3.71 GB
Files
15
Last updated
Jul 7
Pre-warmed CDN
US EU US EU

Contributors